Implementation:Facebookresearch Audiocraft PesqMetric

Knowledge Sources	Facebookresearch_Audiocraft
Domains	Audio_Metrics, Speech_Quality
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for computing PESQ (Perceptual Evaluation of Speech Quality) scores between generated and reference audio.

Description

PesqMetric wraps the PESQ library as a torchmetrics-compatible metric. It resamples audio to 16 kHz, computes per-sample PESQ scores, and aggregates across the batch. Samples where no speech is detected are silently skipped.

Usage

Import this metric when evaluating audio compression or watermarking quality on speech-like signals.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/metrics/pesq.py
• Lines: 1-50

Signature

class PesqMetric(torchmetrics.Metric):
    def __init__(self, sample_rate: int): ...
    def update(self, preds: torch.Tensor, targets: torch.Tensor): ...
    def compute(self) -> torch.Tensor: ...

Import

from audiocraft.metrics.pesq import PesqMetric

I/O Contract

Inputs

Name	Type	Required	Description
preds	torch.Tensor	Yes	Predicted audio [B, C, T]
targets	torch.Tensor	Yes	Reference audio [B, C, T]

Outputs

Name	Type	Description
pesq_score	torch.Tensor	Mean PESQ score

Related Pages

• Principle:Facebookresearch_Audiocraft_Compression_Quality_Evaluation